feat(stats): add Savitzky-Golay smoothing filter

[Shizoqua]

Adds stats.SavitzkyGolay, a rolling Savitzky-Golay smoothing filter as requested in #1424.

Savitzky-Golay fits a polynomial of a given degree to a sliding window of observations by least squares and returns the smoothed value of the most recent point via the precomputed scipy.signal.savgol_coeffs dot product. Because the polynomial is evaluated at the last point of the window, the filter is causal and compatible with online learning.

>>> from river import stats
>>> stat = stats.SavitzkyGolay(window_size=5, polyorder=2)

• Coefficients computed once in __init__, applied over a fixed-size deque
• get() returns None until the window has enough observations
• window_size/polyorder follow scipy's constraint (polyorder < window_size, enforced by scipy)
• Registered in stats/__init__.py and the docs nav
• Release-note entry and CodSpeed benchmark included

Verification:

• pytest tests/stats — 175 passed (incl. doctests)
• Matches a manual least-squares polynomial fit to the same window; reproduces a degree-≤2 polynomial exactly
• ruff check / ruff format --check clean

DCO sign-off included.

[codspeed-hq]

Merging this PR will not alter performance

✅ 98 untouched benchmarks
🆕 1 new benchmark
⏩ 16 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
🆕	Simulation	test_savitzky_golay_update	N/A	5.1 ms	N/A

---

_{Comparing Shizoqua:feat/1424-savitzky-golay (23f2f93) with main (fe248b4)}

Footnotes

1. 16 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[MaxHalford]

Overall looks good, but I would like to see a use case so users can understand when to use this

[MaxHalford]

Is this a list or a numpy array? I think it should be the former

[Shizoqua]

Thanks @MaxHalford! I added a use case to the docstring: a noisy temperature probe drifting upward, where smoothing reveals the underlying trend so it can be fed to a model or a threshold. New commit 7121884 includes the example and the doctest output.

[e10e3]

A commenter in the issue you reference mentioned the article Why and how Savitzky–Golay filters should be replaced, which argues that one should replace the SG filter with other, more stable and less noisy, filters like sinc or Whittaker–Henderson. Is it not better to implement these alternative filters rather than Savitzky–Golay?

Also, did you read the Contributing guide? Your PR cover letter makes me think you did not, but I could be wrong.